A Simple Method to Allow for Guanine-Cytosine Amplification Error

Clinica Chimica Acta 496 (2019) 13–17

Contents lists available at ScienceDirect

Clinica Chimica Acta

journal homepage: www.elsevier.com/locate/cca

A simple method to allow for guanine-cytosine ampliﬁcation error in prenatal DNA screening for trisomy 18 T ⁎ Nicholas J. Wald ,1, Jonathan P. Bestwick1, King Wai Lau, Wayne J. Huttly, Weilin Ke, Ray Cheng, Robert W. Old

Wolfson Institute of Preventive Medicine, Barts and The London School of Medicine and Dentistry, Queen Mary University of London, Charterhouse Square, London EC1M 6BQ, UK

ARTICLE INFO ABSTRACT

Keywords: Background: A source of error in prenatal screening for trisomies is PCR amplification error associated with Prenatal screening guanine-cytosine (GC) content of DNA fragments in maternal plasma. We describe a simple method of allowing Trisomy 18 for this. Edwards syndrome Methods: Data from a Reflex DNA screening programme (67 trisomy 18 and 83 unaffected pregnancies) were Cell-free DNA used to compare the ratio of chromosome 18 DNA fragment counts to chromosome 8 DNA fragment counts GC content (because chromosome 8 has a similar GC content to chromosome 18) with the percentage of chromosome 18 DNA counts using counts from all autosomes in the denominator, with and without an all autosome correction for the GC content of the DNA fragments. Results: A chromosome 18 to 8 ratio of DNA fragment counts was more discriminatory than the percentage of all autosome counts arising from chromosome 18 without, or with an all autosome correction for GC content bias. It achieves a high screening performance, eg. for a 0.25% false-positive rate, a 97% detection rate instead of 49% without a correction for GC content, and 91% with an all autosome correction for GC content. Conclusion: Consideration can be given to using the ratio of chromosome 18 DNA fragment counts to chromosome 8 DNA fragment counts in cell-free DNA prenatal screening for trisomy 18, avoiding the need for more complex methods of making a correction for the GC content currently used.

1. Introduction sequenced DNA fragments from all chromosomes against the GC content of the fragments [4]. Ideally, there should be no association be- Analysis of maternal plasma DNA (also known as cell-free DNA) is tween the GC content of a fragment and the fragment counts sequenced an accurate method for prenatal screening for fetal trisomies 21, 18, so that the plot is horizontal. In practice, however, the plot is bell- and 13 [1]. However the screening performance for trisomy 18, is less shaped, indicating underestimation with DNA fragments with high and than for trisomy 21 [2], and the reasons for this are unknown. This low GC content and overestimation in between. Deviations from the prompted us to examine possible sources of analytical error that might overall average (ie. expectation) can be used to standardize (ie. correct) aﬀect DNA screening for trisomy 18. the error. The method has the advantage of generalizability (eg. ap- The DNA analysis most widely used in prenatal screening for plicable to DNA fragments from all chromosomes) but it has several trisomy 18 is massively parallel sequencing. This involves sequencing disadvantages. The method is prone to variation from analytical run to several million DNA fragments in maternal plasma and then calculating run, and corrections, vary according to the pre-sequencing steps (eg. the proportion of sequences that map to chromosome 18. The de- how the PCR is performed), and according to the sequencing methods nominator of the proportion is usually the number of DNA fragments used, all of which impair analytical precision. This all autosome GC that map to all autosomes. A correction for GC (guanine-cytosine) correction method is complex, not transparent, and requires a large content of the DNA fragments is usually applied [3] to allow for GC dataset, preferably linked to a particular sequencing method and la- associated error in the PCR copying number of DNA fragments. The boratory. usual method for allowing for GC error relies on a plot of the number of Sehnert and colleagues [5] indicate that it may be better to use a

⁎ Corresponding author. E-mail address: [email protected] (N.J. Wald). 1 Joint first authors. https://doi.org/10.1016/j.cca.2019.06.015 Received 25 March 2019; Received in revised form 5 June 2019 Available online 15 June 2019 0009-8981/ © 2019 Published by Elsevier B.V. N.J. Wald, et al. Clinica Chimica Acta 496 (2019) 13–17 single or a small number of chromosomes in the denominator, instead illustrate the fit of chromosome 18 MoM values (using all autosomes as of all autosomes when calculating the proportion of DNA fragments the denominator and using chromosome 8 as the denominator) to aligning to chromosome 18. Empirical testing of different chromosome Gaussian distributions, probability plots were generated separately for denominators indicated that chromosome 8 was the most dis- affected and unaffected pregnancies, with affected pregnancies stan- criminatory for trisomy 18. We explored this strategy as a way of im- dardised to a fetal fraction of 6% by adjusting the MoM values ac- proving DNA screening performance for trisomy 18, using a larger data cording to the slopes of the regression lines. Points on the probability set obtained from the Wolfson Institute (London) prenatal screening plots lying on a straight line indicate a good fit to a Gaussian dis- programme for trisomy 21, 18, and 13 from 2015 to 2018. tribution. The standard deviations of MoM values in affected and unaffected pregnancies were taken as the slopes of regression lines of the points on each probability plot between the 10th and 90th centiles; a 2. Methods standard method of estimating the standard deviation that avoids the undue influence of outliers. [7] The risk of each of the 67 affected and Maternal plasma DNA from 67 trisomy 18 (affected) pregnancies 83 unaffected pregnancies being affected with trisomy 18 was esti- ff and 83 una ected pregnancies was sequenced using a semiconductor mated as the maternal age-specific odds of an affected livebirth [8], sequencing platform and software [6]. Typically about 10 million DNA adjusted to the first trimester by the fetal loss rate from this time in fragments were analysed in each plasma sample. Data from the BAM pregnancy until term [9], multiplied by the likelihood ratio (the height (Binary Alignment Map) files that plasma DNA analysis generated for of the fetal-fraction specific Gaussian distribution in affected pregnan- each pregnancy were aligned to the human reference genome (hg19). cies divided by the height of the Gaussian distribution in unaffected DNA fragments that mapped to individual chromosomes were counted. pregnancies). Screening performance was estimated as the detection The fetal fraction of individual samples was estimated by proprietary rate (DR; the proportion of affected pregnancies with a positive result) (Premaitha) software within the test platform. for a specified false-positive rate (FPR; the proportion of unaffected DNA fragments from chromosome 18 were expressed as (i) a per- pregnancies with a positive result), (FPR for a specified DR and DR and centage using counts from all autosomes as the denominator, without a FPR for a specified risk cut-off). Modelling based on multivariate correction for GC content, (ii) a percentage using counts from all au- Gaussian analyses provides a more robust estimate of screening per- tosomes as the denominator, with an all autosome correction for GC formance, provided the underlying distribution of the markers are ap- content and (ii) a ratio using counts from chromosome 8 as the de- proximately Gaussian, a method that is routinely used in prenatal nominator. Whether a percentage or ratio is used with all autosomes as screening and has been empirically validated. [10–12] It avoids random ff the denominator will make little di erence because the proportion of error in estimation that arises from using directly observed results when DNA fragments from chromosome 18 is small compared to fragments the number of observations is not large, for example if there were no from all autosomes. However, when only considering chromosomes 18 false-positives among 100 unaffected pregnancies in a study sample, it and 8 it does matter because the random error arising from combining should not be taken to mean that there will be no false-positives in the the DNA fragment counts from chromosomes 18 and 8 in the denomi- population at large. We provide observed and modelled results so that a nator is greater than with fragments from chromosome 8 alone. Dot comparison can be made between the two methods. Screening perfor- plots, that categorise observations in a manner that avoids overlapping mance using modelled DNA counts was estimated by simulating data on ff dots, were used to visually compare the distributions in a ected and 100,000 aff ected and unaffected pregnancies based on (i) the dis- unaffected pregnancies with DNA fragments from chromosome 18 ex- tribution of live births in England and Wales 2014–16 [13], (ii) the pressed as (i) a percentage using counts from all autosomes as the de- distributions of fetal fraction in affected and unaffected pregnancies nominator, without a correction for GC content, (ii) a percentage using and (iii) the fetal-fraction-specific Gaussian distributions of chromo- counts from all autosomes as the denominator, with an all autosome some 18 MoM values using (a) all autosomes as the denominator, with correction for GC content and (ii) a ratio using counts from chromo- and without an all autosome correction for GC content, and (b) chro- some 8 as the denominator. mosome 8 as the denominator. The likelihood ratio and risk of being The percentages and ratio were converted into multiple of the affected was then calculated as with the observed data from the 67 median (MoM) values by dividing by the respective median percentages affected and 83 unaffected pregnancies i.e. using observed DNA counts. and ratio in unaffected pregnancies. A regression of log10 MoM values against fetal fraction in affected pregnancies was performed to estimate the fetal-fraction specific median MoM in affected pregnancies. To

Fig. 1. Percentage of DNA fragments in maternal plasma from chromosome 18 (expressed as a percentage of all autosomes) without (a) and with (b) an all autosome correction for GC content and (c) ratio of DNA fragment counts in maternal plasma from chromosome 18 to fragment counts from chromosome 8 in 67 aﬀected and 83 unaﬀected pregnancies.

14 N.J. Wald, et al. Clinica Chimica Acta 496 (2019) 13–17

Fig. 2. Relative frequency distributions of DNA fragment counts in maternal plasma from chromosome 18 expressed as the percentage of all autosome fragment counts from chromosome 18 without (a) and with (b) an all autosome correction for GC content and (c) the ratio of chromosome 18 to chromosome 8 counts (all standardised to a fetal fraction of 6%; in multiples of the unaﬀected median [MoM] values; on a log scale).

3. Results fragment counts from chromosome 18 to chromosome 8. The range of values in unaffected pregnancies is further reduced and there is little Fig. 1a shows the percentage of chromosome 18 DNA fragments in overlap between the values for affected (median 0.535) and unaffected maternal plasma (expressed as a percentage of fragments from all au- pregnancies (median 0.517). tosomes) without any correction for GC content based on 67 affected Appendix Fig. 1 shows the relationship between percent chromo- and 83 unaffected pregnancies. The figure shows higher values in af- some 18 expressed as MoM values using all autosomes as the denomi- fected pregnancies (median 2.87%) than in unaffected pregnancies nator without a correction for GC content in affected pregnancies ac- (median 2.78%), but there is considerable overlap in values. Fig. 1b cording to fetal fraction together with a regression line. MoM values shows the same, but with an all autosome correction for GC content. increased by 0.26% for each percentage point increase in fetal fraction The range of values in unaffected pregnancies is much reduced and (p < 0.001). Appendix Fig. 2 shows the same, but with an all autosome there is little overlap between the values for affected (median 3.21%) correction for GC content. MoM values increased by 0.32% for each and unaffected pregnancies (median 3.10%). Fig. 1c shows the ratio of percentage point increase in fetal fraction (p < 0.001). Appendix Fig. 3

15 N.J. Wald, et al. Clinica Chimica Acta 496 (2019) 13–17

shows the same, but using the ratio of chromosome 18 to chromosome 8 counts. MoM values increased by 0.33% for each percentage point increase in fetal fraction (p < 0.001). The regression line equations used to define the fetal fraction-specific median MoM in affected pregnancies, and the standard deviations of MoM values in affected pregnancies

are given in Appendix Table 1. The standard deviation of log10 MoM 97 0 1 in 150 95 0 values in unaffected pregnancies was statistically significantly smaller for the ratio of fragment counts from chromosome 18 to chromosome 8 than the percentage of chromosome 18 DNA fragments using all autosomes as the denominator with an all autosome correction for GC content (p = 0.002). Appendix Figs. 4, 5 and 6 show probability plots of MoM values using chromosome 18 counts as a percentage with an all autosome 97 0 93 0 1 in 100 denominator without and with correction for GC content and using the ratio of chromosome 18 to chromosome 8 counts respectively, with MoM values in affected pregnancies standardised to a fetal fraction of 6%. All the plots are reasonably linear indicating a good fit to Gaussian distributions. The plots for affected and unaffected pregnancies in Appendix Fig. 6 are “flatter”, indicating smaller standard deviations using the ratio of chromosome 18 to chromosome 8 counts than chro- and expressing chromosome 18 (chr18) DNA fragments counts as the 97 0 91 0 60 1.2 64 1.2 73 1.2 1 in 50 mosome 18 counts as a percentage with an all autosomal denominator ff either with or without a correction for GC content. Fig. 2 shows the modelled distributions (using medians and standard deviations derived from Appendix Figs. 4, 5 and 6) using chromosome 18 counts as a percentage with an all autosome denominator without (Fig. 2a) and with (Fig. 2b) an all autosome correction for GC content and using the ratio of chromosome 18 to chromosome 8 counts (Fig. 2c), with MoM values in affected pregnancies standardised to a 97 0 91 0 55 0 1 in 40 fetal fraction of 6%. The figure shows greater discrimination (Fig. 2c) using the ratio of chromosome 18 to chromosome 8 which arises from the smaller standard deviation in unaffected pregnancies and hence reduced overlap between the distributions in affected and unaffected pregnancies. Table 1 shows the trisomy 18 detection rates and false-positive rates based directly on the observed data and based on the modelled results 97 0 90 0 52 0 1 in 30 using the distribution parameters in Appendix Table 1. The observed and modelled results are similar, for example at a 1 in 50 risk cut-off, using the ratio of chromosome 18 to chromosome 8 counts the detection rate is 97% (65/67) and false-positive rate is 0% (0/83) compared with the modelled estimates of 95% and 0.05% respectively. There is a clear improvement in performance when the ratio of chromosome 18 to chromosome 8 counts is used instead of using chromosome 18 counts as of:- 97 0 88 0 49 0 1 in 20 a percentage with an all autosome denominator either with or without ff an all autosome correction for GC content. For example, at a risk cut-off of 1 in 50 the DR is 95% and the FPR 0.05% using the ratio of chromosome 18 to chromosome 8 counts compared with a DR of 89% and an FPR of 0.12% using chromosome 18 counts as a percentage with an all autosomal denominator with correction for GC content and a DR of 57% and an FPR of 0.53% using chromosome 18 counts as a percentage with an all autosomal denominator without correction for GC content. 94 0.01 95 0.02 95 0.03 95 0.04 95 0.05 96 0.1 96 0.15 97 0 38 0.04 46 0.18 51 0.30 55 0.43 57 0.53 66 1.1 71 1.6 39 0 DR and FPR for risk cut-o DR (%) FPR (%) DR (%) FPR (%) DR (%) FPR (%) DR (%) FPR (%) DR (%) FPR (%) DR (%) FPR (%) DR (%) FPR (%) 1 in 10 Table 2 shows the detection rates for specified false-positive rates and false-positive rates for specified detection rates using the modelled data. For example, for a false-positive rate of 0.25% the use of the ratio of chromosome 18 to chromosome 8 counts yields a 97% detection rate compared with 91% and 49% using chromosome 18 counts as a percentage with an all autosome denominator with and without an all autosome correction for GC content respectively.

4. Discussion

Our results show that in prenatal screening for trisomy 18 the ratio of plasma DNA fragment counts that map to chromosome 18 to DNA fragment counts that map to chromosome 8 is a simple method of allowing for analytical error due to variation in DNA GC content. It is With all autosome correction for GC content 85 0.03 87 0.06 88 0.08 89 0.1 89 0.12 91 0.25 92 0.36 Without correction for GC content Without correction for GC content With all autosome correction for GC content 84 0 Ratio chr18 to chr8 fragment counts % of all autosome fragment counts from chr18 % of all autosome fragment counts from chr18 Ratio chr18 to chr8 fragment counts clearly better than making no adjustment for GC content and our results Modelled DNA counts Observed DNA counts Chromosome 18 MoM Table 1 Detection rates (DRs) and false-positive rates (FPRs) using observed DNA fragment counts and modelled counts according to trisomy 18 risk cut-o percentage of counts from all autosomes without and with all autosome correction for GC content or as the ratio to chromosome 8 (chr8) fragment counts. indicate that the method is also better than the conventional method of

16 N.J. Wald, et al. Clinica Chimica Acta 496 (2019) 13–17

Table 2 Detection rates (DRs) according to false-positive rates (FPRs) and FPRs according to DRs using modelled DNA fragment counts according to expressing chromosome 18 (chr18) DNA fragments counts as the percentage of counts from all autosomes without and with all autosome correction for GC content or as the ratio to chromosome 8 (chr8) fragment counts.

Chromosome 18 MoM DR (%) for FPR of:- FPR (%) for DR of:-

0.125% 0.25% 0.50% 1% 2% 85% 90% 95% 98% 99%

% of all autosome fragment counts from chr18 Without correction for GC content 42 49 56 64 73 6.1 10 19 32 41 With all autosome correction for GC content 89 91 93 95 96 0.03 0.15 1.1 5.7 12 Ratio chr18 to chr8 fragment counts 96 97 97 98 99 < 0.01 < 0.01 0.03 0.96 3.8 all autosome GC adjustment. Sehnert et al. [5] observed such an ad- Acknowledgements vantage without linking it to DNA fragment GC content. Our data take the observation further by estimating the quantitative effect using the We thank Tiesheng Wu for providing IT support and help. ratio of chromosome 18 to chromosome 8 fragment counts without GC correction and finding an improved screening performance; the im- Appendix A. Supplementary data provement in performance is considerable. Chromosome 8 has a GC- content that is close to that of chromosome 18 [14]. Using the ratio of Supplementary data to this article can be found online at https:// chromosome 18 to chromosome 8 fragment counts therefore directly doi.org/10.1016/j.cca.2019.06.015. allows for variability in the PCR part of the DNA analysis arising from GC associated error in amplifying the correct number of DNA fragments References which can vary from sample to sample, which affect DNA fragment counts. Failure to take account of the GC content of DNA has a clinically [1] G.E. Palomaki, C. Deciu, E.M. Kloza, G.M. Lambert-Messerlian, J.E. Haddow, significant effect on screening performance as illustrated in Tables 1 L.M. Neveux, M. Ehrich, D. van den Boom, A.T. Bombard, W.W. Grody, S.F. Nelson, J.A. Canick, DNA sequencing of maternal plasma reliably identifies trisomy 18 and and 2. trisomy 13 as well as down syndrome: an international collaborative study, Genet. The conventional method of allowing for GC content is based on Med. 14 (3) (2012) 296–305. algorithms using estimated associations between the number of DNA [2] M.M. Gil, V. Accurti, B. Santacruz, M.N. Plana, K.H. Nicolaides, Analysis of cell-free DNA in maternal blood in screening for aneuploidies: updated meta-analysis, fragments and GC content of those fragments. The method has the Ultrasound Obstet. Gynecol. 50 (3) (2017) 302–314. limitation of its complexity and a lack of transparency. The algorithm [3] E.Z. Chen, R.W.K. Chiu, H. Sun, R. Akolekar, K.C. Allen Chan, T.Y. Leung, et al., adopted has to be derived separately for each method of DNA analysis. Noninvasive prenatal diagnosis of fetal trisomy 18 and trisomy 13 by maternal The method proposed here is simple, transparent, and does not require plasma DNA sequencing, PLoS ONE 6 (2011) e21791. [4] Y. Benjamini, T.P. Speed, Summarizing and correcting the GC content bias in high- the use of an algorithm. The method is also generalizable in that any throughput sequencing, Nucleic Acids Res. 40 (10) (2012) E72. chromosome with a similar GC content to the trisomic chromosome of [5] A.J. Sehnert, B. Rhees, D. Comstock, E. de Feo, G. Heilek, J. Burke, R.P. Rava, interest could be used as the reference, using chromosome 7 in Optimal detection of fetal chromosomal abnormalities by massively parallel DNA sequencing of cell-free DNA in maternal blood, Clin. Chem. 57 (7) (2011) screening for trisomy 21 and chromosome 4 in screening for trisomy 13. 1042–1049. These would need to be validated empirically as we have done here [6] F. Crea, M. Forman, R. Hulme, R.W. Old, D. Ryan, R. Mazey, M.D. Risley, The Iona with trisomy 18 screening, before being adopted in practice. test: development of an automated cell-free DNA-based screening test for fetal trisomies 21, 18 and 13 that employs the ion proton sequencing platform, Fetal Diag. Truncation limits that limit the range of values used to calculate Ther. 42 (3) (2017) 218–224. likelihood ratios and hence risk are often used for markers in prenatal [7] N.J. Wald, H.S. Cuckle, J.W. Densem, K. Nanchalal, P. Royston, T. Chard, screening. However, when the distributions of marker values in affected J.E. Haddow, G.J. Knight, G.E. Palomaki, J.A. Canick, Maternal serum screening for – ff Down's syndrome in early pregnancy, BMJ 297 (1988) 883 887. and una ected pregnancies are widely separated, as is the case here, [8] G.M. Savva, K. Walker, J.K. Morris, The maternal age-specific live birth prevalence this has the effect of ignoring many of the informative values. In such of trisomies 13 and 18 compared to trisomy 21 (down syndrome), Prenat. Diagn. 30 situations not using truncation limits is appropriate provided the (2010) 57–64. [9] J.K. Morris, G.M. Savva, The risk of fetal loss following a prenatal diagnosis of modelled estimates are consistent with those based on simple counting trisomy 13 or trisomy 18, Am. J. Med. Genet. Part A 146A (2008) 827–832. (see Table 1). While the precise risk at high MoM values, for ex- [10] N.J. Wald, J.P. Bestwick, W.J. Huttly, Improvements in antenatal screening for ample > 1.05 in Appendix Fig. 6, may be uncertain, the probability of Down's syndrome, J. Med. Screen. 20 (2013) 7–14. being affected becomes extremely high and has little influence on [11] J.P. Bestwick, W.J. Huttly, N.J. Wald, Detection of trisomy 18 and trisomy 13 using first and second trimester Down's syndrome screening markers, J. Med. Screen. 20 screening performance. Similarly, at low values, for example < 0.98 the (2013) 57–65 (Corrigendum in J Med Screen 2015;22:52–4). pregnancy will almost certainly be unaffected. [12] N.J. Wald, W.J. Huttly, K.W. Murphy, K. Ali, J.P. Bestwick, C.H. Rodeck, Antenatal In summary, the use of the ratio of DNA fragment counts that map to screening for Down's syndrome using the integrated test at two London hospitals, J. Med. Screen. 16 (2009) 7–10. chromosome 18 to counts that map to chromosome 8 yields a high level [13] Office for National Statistics, Birth Characteristics in England and Wales, 2015 of screening performance and avoids the need for more complex GC (2014), p. 106. correction algorithms. [14] http://blog.kokocinski.net/index.php/gc-content-of-human-chromosomes? blog=2.

Declaration of Competing Interest

The authors have no interests to declare.